Im Rahmen des PRIME (Paderborn Research Colloquium on Information Management & Engineering) begrüßt das Department Wirtschaftsinformatik Prof. Dr. Martin Bichler (TUM). Wir laden alle Interessierten zu seinem Vortrag "Revenue in First- and Second-Price Display Advertising Auctions: Understanding Markets with Learning Agents" ein.
Martin Bichler is head of a research unit and a full professor at the Department of Computer Science of the Technical University of Munich (TUM). He is also affiliated with the TUM School of Management and a core member of the Munich Data Science Institute. Martin received his MSc degree from the Technical University of Vienna, and his Ph. D. as well as his Habilitation from the Vienna University of Economics and Business. He was a research fellow at UC Berkeley, and a research staff member at the IBM T. J. Watson Research Center, Yorktown Heights, New York. Later, he was a visiting scholar at the University of Cambridge, at HP Labs Palo Alto, at the Department of Economics at Yale University, and at the Department of Economics at Stanford University. In the fall 2023, he was a Research Professor at the Simons Laufer Mathematical Sciences Institute . in Berkeley.
Abstract: The transition of display ad exchanges from second-price to first-price auctions has raised questions about its impact on revenue. Auction theory predicts the revenue equivalence between these two auction formats. However, display ad auctions are different from standard models in auction theory in at least two important ways. First, automated bidding agents cannot easily derive equilibrium strategies in first-price auctions because distributional information regarding competitors' values or even the number of competitors is not readily available. Second, due to principal-agent problems, bidding agents typically maximize return on investment (ROI) and not payoff. The literature on learning agents for real-time bidding in these auctions is growing because of the practical relevance of this area. However, whether such learning agents converge to an equilibrium is an open question. Learning dynamics can cycle or even be chaotic. Recent experiments suggest that they might also end up in collusive low-price outcomes. Whether bids are in equilibrium or not cannot easily be determined from field data since bidders' underlying values are unknown. In this paper, we derive equilibrium predictions and explore the convergence behavior of widespread online learning algorithms. In addition, we leverage recent developments in equilibrium computation in order to get equilibrium predictions in situations where analytical solutions to the governing differential equations cannot be derived. Contrary to prior results, we show that widespread learning algorithms do not exhibit systematic deviations and instead learn to play according to equilibrium even if bidders maximize ROI and not payoff. Our results show that collusion may not be the explanation for lower revenue with first-price auctions. Instead, we show that even in equilibrium, second-price auction achieves higher expected revenue compared to the first-price auction with ROI-maximizing bidders, i.e., the change in auction format might have had substantial and non-obvious consequences for ad exchanges and advertisers bidding for display ads.
